Interview
[Noon Salon] Semicon Research Lab CEO Roh Geun-chang: AI-Driven Memory Demand Cycle Still in Early Stage
Today's Noon Salon featured Roh Geun-chang, CEO of Semicon Research Lab, who spent 28 years analyzing the semiconductor industry as an analyst, served 13 years as a research center head, and recently went independent to found his own firm. Regarding recent remarks by Anthropic CEO Dario Amodei and others warning of AI's risks, Roh interpreted them—coming ahead of an IPO—as serving two purposes: raising the company's valuation multiple by conveying the message that AI is capable enough to replace human labor, while also raising entry barriers by emphasizing compliance and safety, which increases the cost burden on later-stage startups founded by people from OpenAI, Anthropic, and Google.
Roh dismissed concerns that this call for a measured pace would lead to reduced AI investment or scaled-back data centers. He explained that while adding security and verification functions to chip design raises costs for both foundries and memory makers, this cost increase ultimately translates into price premiums, becoming a factor that improves semiconductor companies' profitability. He cited TSMC's stock, which barely declined even amid this week's market correction, as an example. He also noted that consensus estimates for next-quarter operating profit at Samsung Electronics and SK Hynix keep being revised upward, with Samsung Electronics projected to exceed roughly 14 trillion won and SK Hynix to reach approximately 7-8 trillion won.
He noted it is industry consensus that U.S. data center power demand will triple from the current roughly 40GW to 120GW by 2030, and that according to TrendForce, the HBM market is projected to roughly double from about $60 billion this year to $120 billion next year, and reach $560 billion by 2030. Google's external TPU sales volume is expected to reach 4.5 million units by 2028, versus 12 million units for Nvidia over the same period, and Google is reportedly set to become the world's first to adopt HBM4E next year. Given this volume expansion and improving HBM4 mix, he projected that despite concerns over won strength, both Samsung Electronics and SK Hynix will see revenue exceed market consensus, with operating profit also likely to come in better than the lower end of current worst-case expectations.
He explained that SK Hynix's HBM market share could decline from about 59% this year to around 52% as Samsung Electronics catches up, but since the HBM market itself is expanding, SK Hynix's HBM revenue itself will actually increase. Regarding Samsung Foundry, he pointed out that it has yet to secure major customers in the data center and AI accelerator segment as a weakness, contrasting this with TSMC, where data centers account for over 60% of revenue. However, he added that since Samsung Electronics and TSMC are the only companies with EUV equipment, its competitiveness in fine process technology remains valid, and it needs to gradually expand its presence by securing customers such as Tesla in the automotive sector.
On reports of a potential joint production review between SK Hynix and Intel in the U.S., he assessed that since Intel continues to post losses in advanced foundry and faces a shortage of products to allocate to facilities like its Arizona plant—as more customers in the AI PC and server CPU markets shift to their own ARM chips—discussions of a joint venture to utilize the site are plausible. However, he noted that concerns over memory technology leakage mean a cautious approach is needed. Finally, on semiconductor sector valuations, he said it is rare for a sector to satisfy all three conditions of industry growth potential, undervalued valuation, and high dividend yield simultaneously, projecting that if Samsung Electronics' annual dividend of roughly 14,000 won is realized, its dividend yield could reach 6%. He noted that as recent rate hikes have widened valuation discounts for U.S. hardware companies, Korean semiconductor stocks have also entered a relatively more undervalued phase.